Revisiting data weighting in fisheries stock assessment models

This paper revisits topics addressed in two previous papers on data weighting in fisheries stock assessment models: the first was general (Francis, 2011; Can. J. Fish. Aquat. Sci. 68, 1124-1138); the second considered the related problem of finding the best likelihood for composition data (Francis, 2014; Fish. Res. 151, 70-84). In the light of subsequent literature and experience, four topics seemed in need of increased emphasis or elaboration. (1) For composition data, it is better to think in terms of “right-weighting” (i.e., weighting that is statistically appropriate) than “down-weighting”. (2) The sensitivity of some assessments to changes in weighting can sometimes be reduced by restructuring to reduce model misspecification; this is a good idea, but should be seen as complementary to data weighting, rather than an alternative to it. (3) It seems typical that more than half of the variance of composition residuals arises from process error (arising from model misspecification) rather than observation error. (4) Changing the likelihood for composition data from the multinomial to the Dirichlet-multinomial has some advantages but is not without problems. Some new topics are discussed: most iterative reweighting of composition data is multiplicative, but additive methods deserve consideration; data weighting is more complicated in state-space models; catch data should not be subject to data weighting; there are significant disadvantages in structuring age-related observations of fishery and survey catches as frequencies, rather than compositions (i.e., as numbers, rather than proportions); methods of weighting three additional data types (conditional age at length, tagging abundance; and tagging length-increment) are described.